VibeVoice on a Mac: What It Does and What to Run Instead

Short answer: Microsoft's VibeVoice is a research text-to-speech family that can synthesize up to 90 minutes of conversational audio with up to four distinct speakers. It is not a Mac app, the current documentation has no supported Apple Silicon install path, and Murmur does not run it. Microsoft removed the VibeVoice-TTS inference code in September 2025, so for most Mac users the realistic choice is a shipped local model rather than a reproduction of the demo.

The weights are still on Hugging Face, and the related Realtime 0.5B and ASR models remain available. That split is the source of most of the confusion around the name. This guide separates what Microsoft still publishes from what it has withdrawn, then compares the model with a Mac workflow for multi-voice scripts.

What VibeVoice was built to do

VibeVoice-TTS turns text into conversational speech using a next-token diffusion design. A Qwen2.5-1.5B language model reads the script and the dialogue structure, acoustic and semantic tokenizers compress audio at a 7.5 Hz frame rate, and a diffusion head reconstructs the acoustic detail. Microsoft trained the released 1.5B model with a context curriculum that rises to 65,536 tokens.

Two details matter more than the headline. The 1.5B name refers to the language model, not the whole repository: Hugging Face shows about 2.7B parameters for the repository, more than the 1.5B in the name. And the card is explicit that the model is speech only and does not model crosstalk. A podcast with people talking over each other is not what this model produces, even though its demo page shows lively conversation.

The TTS code is disabled, and that changes the decision

Microsoft released VibeVoice-TTS on 2025-08-25. On 2025-09-05 the team wrote that it had found uses inconsistent with the stated intent and removed the VibeVoice-TTS code from the repository. The documentation page still exists, but its Installation and Usage section now reads, in full, "Disabled due to widespread misuse." There are no supported install commands to copy. The project page repeats that the repository was disabled until Microsoft is confident that out-of-scope use is no longer possible.

The model table marks VibeVoice-TTS-1.5B as "Disabled" in the Quick Try column, and VibeVoice-Large is marked Disabled as well. The Large model card returned an authorization error when I opened it. The 1.5B weights do still resolve on Hugging Face, where the file listing shows 5.41 GB and the page reported roughly 537,000 downloads in the last month at the time of checking. Downloadable weights are not the same thing as a supported pipeline, and that distinction is the whole story here.

The Realtime 0.5B model is a separate release and is still documented. It is single speaker, built for streaming text input, and generates about ten minutes with an 8k context window. It is not a substitute for the 90 minute multi-speaker job, and its own install guide points at an NVIDIA container.

One naming collision is worth clearing up once. vibevoice.net is a dictation and speech-to-text product with Windows, macOS, and Linux clients. It is not Microsoft's model and it does not generate speech. If you arrived from a search result about typing by voice, you are looking at a different product with a shared name.

What it actually takes to run on a Mac

Microsoft does not publish a Mac hardware table for VibeVoice-TTS. The only installation guidance left in the repository is for the Realtime model, and it recommends an NVIDIA PyTorch container with GPU access plus a manual FlashAttention install. The same Realtime document also says an NVIDIA T4 and a Mac M4 Pro achieve real-time performance in Microsoft's tests, which is the one official Apple data point. That is a Linux and CUDA path, and it is the only path the official docs describe.

The issue tracker fills part of the gap, with the caveat that these are user reports rather than requirements. In issue #26, a commenter says the 1.5B model runs on an M1 with 32 GB of memory but at low tokens per second. In issue #119, one user reports realtime output on an M1 Max with 64 GB that runs but sounds choppy, while a Microsoft collaborator replies that the code runs well on an M4 and that playback smoothness varies by hardware while the saved audio is correct. Another user in the same thread reports crackling on an M4 with 24 GB during realtime playback but clean downloaded files.

One report concerns the 1.5B model and the other the Realtime model, and both are community reports, not official guidance. The 1.5B TTS model carries more parameters and a 64K context, so treat any realtime figure as a floor rather than a promise. A separate MPS and CPU crash in streaming inference, caused by a mask size bug under SDPA attention, was reported in issue #312 and fixed in April 2026. If you install from an older commit, that failure is part of your setup.

Disk use is the predictable part. The 1.5B weights are 5.41 GB before you add a Python environment, PyTorch, Transformers, and any duplicate downloads in a model cache. Leave room for those, and expect the first run to spend time loading rather than generating.

License and the responsible-use terms Microsoft attaches

The repository and the model card are MIT licensed, which is permissive. That is not the whole agreement. Microsoft attaches research-only guidance and a list of out-of-scope uses to the release, and those terms are specific.

• Voice impersonation without explicit, recorded consent, including satire, advertising, ransom, social engineering, or authentication bypass.
• Disinformation or audio presented as a genuine recording of a real person or event.
• Real-time or low-latency voice conversion, including telephone and video-conference use.
• Unsupported languages: the card says training covered English and Chinese and that other languages may be unintelligible or offensive.
• Background ambience, Foley, or music, because the model is speech only.
• Use to generate any text transcript, or in any manner that violates applicable laws.

The safeguards are unusual enough to plan around. The card says every synthesized file contains an audible disclaimer such as "This segment was generated by AI" plus an imperceptible watermark that lets third parties verify provenance. It also says inference requests are logged in hashed form for abuse detection. For a finished podcast or a client deliverable, an audible AI tag baked into the file is a production constraint, not a footnote.

The README adds that Microsoft does not recommend commercial or real-world use without further testing and development. If your plan is a paid client project, that sentence belongs in your risk assessment, and so does the MIT grant, which does not override the out-of-scope list.

Should you run it yourself?

What can go wrong

• The official TTS install instructions are disabled, so any working setup comes from community forks, older commits, or a third-party wrapper, none of which Microsoft supports.
• FlashAttention is part of the documented GPU path. Issue #312 reports that on CUDA without flash_attn the code falls back to SDPA, so check the current README for what your setup needs.
• MPS and CPU inference hit a mask size crash under SDPA attention until the April 2026 fix, and older environments still carry it.
• Realtime playback is reported as choppy on several Apple machines even when the saved file is correct, which makes live monitoring unreliable.
• The model can add background music or sound effects on its own. The FAQ calls this an intentional easter egg and says it cannot be directly controlled, which is a problem for clean narration.
• Chinese output is less stable than English, and the docs recommend English punctuation to reduce it.
• There is no text normalization, so numbers, symbols, and unusual names depend entirely on the language model.
• VibeVoice-Large is disabled and its model card is not publicly readable, leaving 1.5B as the only reachable multi-speaker checkpoint.

The honest Mac alternative for multi-voice scripts

If the goal is a script with several speakers that you can revise and export, the model choice matters less than the loop around it. Murmur is a local voice studio for Apple Silicon and Intel Macs on macOS 15 or later. Its Projects workspace holds several scripts and a shared speaker roster, and each speaker can use a preset, community, cloned, or designed voice. Generated segments become clips with takes, trims, fades, gain, mute state, markers, and media lanes, and the project exports a full render or separate clips as WAV or M4A.

Murmur runs Kokoro, Chatterbox Turbo, Chatterbox Multilingual, Qwen3-TTS Base, Voice Design, SparkTTS, Fish Audio S2 Pro, Breeze TTS 2, and OmniVoice. It does not run VibeVoice, and it does not do speech-to-text. Models other than bundled Kokoro ask for at least 16 GB of unified memory, according to the app's model definitions, so check that before downloading a large engine. The local studio versus DIY comparison covers the tradeoff between a packaged workflow and assembling your own pipeline.

The two tools solve different problems. VibeVoice's single-pass 90 minute generation is genuinely unusual and hard to match with a clip-based editor. Murmur's advantage is the work around generation: a roster, a queue, alternate takes, and a timeline that exports. If the job is a 90 minute unbroken conversation in one pass, VibeVoice is the specific tool and no Mac app replaces it. If the job is a weekly 20 minute show that gets revised, the editing loop matters more than the single-pass record. The multi-voice podcast workflow walks through the clip-based version.

A two-voice test you can run in an afternoon

Before committing to any model, test the part that is hard to fix later: whether two voices stay distinct across a full script. A short demo sentence will not tell you that.

1. Write 12 short lines that alternate between two speakers. Use roles such as host and guest, not the name of a real person or a copyrighted character.
2. Create a project, add two speakers, and assign a different voice to each. Pick voices that differ in pitch and pace, not just in accent.
3. Paste the lines as separate segments and label each one with its speaker.
4. Generate one line per speaker first and listen back to back. If the two voices blur together, change one before generating the rest.
5. Generate the remaining lines, then regenerate only the lines that mispronounce a word or land with the wrong emphasis.
6. Add markers at segment boundaries so the editor can find each turn without scrubbing.
7. Export WAV for editing, or M4A for direct delivery. Keep the project file so a later correction is one clip, not a full re-render.

If either voice is a clone, use only your own voice or a voice you have written permission to use. Consent should name the project and the period of use, and it should be revocable. The AI voice disclosure guide covers how to label the result on platforms that expect it.

For a wider view of what each local engine does well, the local TTS model roundup compares the shipped options on language coverage, cloning support, and stability. Read it before you download three models to test the same five lines.

Frequently asked questions